DE3043516C2 - Method and device for speech recognition - Google Patents

Abstract

Speech sound recognition is made using a reduced number of speech parameter elements, e.g., five correlation coefficients rather than sixteen spectral coefficients. The five correlation coefficients are derived from comparison of the spectral coefficients of unknown or standard sounds against the spectral coefficients of five highly-separable vowel-like sounds. Then, unknown-sound correlation coefficients are compared with standard-sound coefficients for recognition.

Description

a) Calculating ρ first characteristic parameters from a speech input signal at individual points in time, where ρ is a positive. integer> 2, is

b) Carrying out a comparison operation between the first characteristic parameters calculated in the first step a) and a number of ρ first characteristic parameters of η-specific speech patterns of high distinctiveness at individual points in time to determine a number of η second characteristic parameters that determine the distance of the correlation between the represent the first characteristic parameters of the speech input signal and the first characteristic parameters of the specific speech musician, where η is a positive integer with 1 <n <p ,

c) Determination of a coincidence between the obtained second characteristic parameters of the Mt speech input signal cs and the η second characteristic parameters of each standard language corresponding to Jem method step

b) calculated from the first characteristic parameters of each standard prari - ■ and stored before the speech input signal is supplied.

2. The method according to claim 1, characterized in that that in the first step a) a normalization of the first characteristic parameters is carried out will.

3. The method according to any one of claims 1 or 2, characterized in that for the specific Speech signals any vowels. Rcib consonants and / or nasal consonants are used.

4. The method according to any one of claims 1 to 3, characterized in that in the second step a smoothing of the time sequence data of the second characteristic parameters is carried out and these data

be sampled at predetermined intervals. ; ii

5. Device for performing a method according to one of claims 1 to 4, characterized by a first device (21-26) for calculating ρ first characteristic parameters from speech input signals (1) at corresponding times, where ρ is a positive integer greater than or equal to 2 is a must-crane fitting unit (3). which compares the first characteristic parameters calculated by the first device (21-26) with ρ first characteristic parameters of η given, ni-specific Spraclinuisiern high Undefseheklbm'-wedge yu individual points in time and / while supplies characteristic parameters, where η is a positive whole number 1 </;<:/> is. by a determination unit (4) which determines a coincidence between the wide characteristic parameters of the speech input signal obtained from the misalignment unit (3) and the second characteristic parameters previously determined from each standard language.

6. Apparatus according to claim 5, characterized in that that the first execution (21-26) is a standardization facility (26) for normalizing the first characteristic parameters.

7. Apparatus according to claim 5 or 6, characterized in that the pattern adaptation unit (3) means (242-1,242-16) for smoothing the sequence data of the second characteristic parameters and sampling this data at predetermined intervals.

The invention relates to a method for speech recognition according to the preamble of claim 1 as well as a device for its implementation.

In general, a device for speech recognition can be structurally divided into (1) a computing unit for calculating characteristic parameters of speech and (2) a decision circuit for adapting the characteristic parameters to standard parameters and for determining and recognizing the language. So far, a linear prediction coefficient, a partial autocorrelation coefficient, have been the characteristic parameters of language and a filter bank output signal has been used. If any of these parameters are used, the speech pattern in the generally expressed as a time sequence of parameters of some orders. If an autocorrelation coefficient eighth order, which is sampled every 10 milliseconds, must be taken into account. are for example 800 parameters are required to express the language of a second, so that a large increase in processing is necessary to make a comparison operation with the parameters required for recognition through the standard grid. aside from that a large amount of memory is required to store the standard patterns. Accordingly, ss required to obtain effective characteristic parameters capable of communicating the speech pattern with to represent a possible high fidelity.

From the DLi-OS 23 47 738 is a method for speech recognition known, in which an excerpt from special doctrinal and exercise words is obtained. For Such a feature extract is stored for each individual phrase or practice word. In speech recognition is based on the same rules that were used to obtain the extract of characteristics for the exercise words, a feature excerpt also prepared for the voice input signals. The extract of features of the current one Speech input signal is then compared with the stored feature extracts of the exercise words, and the comparison result is represented by a correlation number. Using the correlation number, then it is determined with what probability the language inuniissipiiiil with your respective (Ib'ingwort aboutcinstitnn'.i. its characteristic extract is saved isi.

In this known method, too, must be within in short, a large number of comparative options be carried out so that the total for the Spracherkennuni; The processing time required becomes large.

Closing the hrfindunu is' therefore, a procedure and

: inc device for speech recognition to indicate with Jem the number of parameters in a comparison siperation are attempted with parameters from standard musicians, in order to save the for to reduce the comparison operation and thus the processing time required for speech recognition and at the same time the storage capacity required for storing the parameters of the standard pattern to decrease.

This object is achieved with a method according to the preamble of claim 1, which according to the invention is designed according to the manner specified in the characterizing part of this claim.

Advantageous embodiments of the method according to the invention and a preferred device for it Implementation are specified in the subclaims.

According to the invention, ρ characteristic parameters calculated from the speech input signals are each compared with ρ characteristic parameters of η specific languages, their mutual distance or the probability of their correspondence being determined. The data obtained in this way are used for 15 seconds characteristic parameters of the input languages and compared with characteristic parameters of standard samples which were determined in the same way. In this way the speech recognition is carried out.

As the η characteristic parameters of the specific language, for example, the characteristic parameters of those vowels which can be clearly distinguished from one another and which have a high degree of separability can be used.

The invention is explained below on the basis of the description of exemplary embodiments and with reference explained in more detail on the accompanying drawing. The drawing shows in

1 shows a block diagram to explain the basic structure of a device according to the invention for speech recognition;

2 shows a block diagram to explain the essential parts of the device according to the invention:

3 shows a block diagram of an exemplary embodiment for a filter bank according to FIG. 2;

Fig. 4 is a diagram for explaining the characteristic Parameters of vowels: and in

Fig. 5 is a block diagram for explaining the normalization process, Music buffer and distance calculation circuit according to Fig. 2.

Fig. 1 shows the basic structure of the novel device for speech, where the reference number I a speech input signal, the reference number 2 an analysis unit, the reference number 3 a pattern matching unit, reference numeral 4 a decision unit, and reference numeral 5 an output signal of the recognition result.

With such a structure, the analyzing unit 2 calculates characteristic parameters from the speech input signal Xp , compares these calculated characteristic parameters with the jcweiligci ρ characteristic parameters of η specific languages and determines such characteristic parameters which represent the distance or the probability between the two types of parameters . The characteristic parameters determined in this way are then compared in the same way with characteristic parameters of the standard pattern, which have been determined in the same way by comparison with characteristic parameters of specific languages, in the pattern adaptation unit 3, and the decision unit 4 recognizes the speech input signal and runs the result of the Detection as output signal 5

Putty. 2 zemi a block diagram of an implementation

> examples for the construction of the essential parts of the novel device for speech recognition, the essential parts corresponding to the analysis unit 2 in FIG.
In Fig. 2, the reference numeral 21 denotes a control

U) circuit, consisting of sequence controllers or the like, which generate various types of control signals. The reference number 22 denotes an analog-to-digital converter, numeral 23 an input buffer memory. the reference number 24 a filter bank, reference numeral 25 an intermediate buffer memory, reference numeral 26 a normalizing circuit, reference numeral 27 a pattern buffer memory, reference numeral 28 a micro-pattern buffer memory. the reference number 29 a distance calculation circuit and the

: n reference numeral 30, a parameter buffer memory.

This embodiment is so blown that it the Aüsgar.gssigruile of the filter bank as characteristic parameters which represent the speech input signal.

The speech input signal applied to the analysis unit 2 is converted into digital signals with the analog-to-digital converter 22 and sent to the input buffer memory 23 is applied, a control being carried out by the control circuit 21. The data sent to the input buffer

Xi memory 23 applied digital signals are in the same Manner subject to the control of the control circuit 21 and applied to the filter bank 24. As in Fig. 3 Shown, this filter bank 24 consists for example of bandpass filters 241-1 to 241-16, the frequency range in the telephone system, i.e. dividing 300 to 3600 Hz logarithmically into 16 equidistant channels, as well as Smoothing and sampling circuits 242-1 through 242-16 for smoothing and re-sampling these output signals the bandpass filter. The 16 smoothed pieces of data.

which are scanned, for example, at 10 second intervals, are transmitted to the intermediate buffer memory 25 as a vector of the characteristic parameters u at the times /. The time sequence of this vector of characteristic parameters a is arranged sequentially in the intermediate buffer memory 25. It is assumed here that S ₁ = ', «,,. </ ,, .... </ _16ι } and α _μ indicates the output signal of the / th filter at time ι. In general, the size of Uj changes with the input pcgcl of the speech input signal 1 and with a

5 (i language environment, so that the comparison of absolute values is difficult. Correspondingly, the vector of characteristic parameters u, which is read out from the buffer memory 25 by the control action of the filter circuit 21, is applied to the normalization circuit 26 in order to to obtain a normalized vector of characteristic 'nramciern a' . The / te F.lemcni of this vector a \ is obtained according to the following equation:

ld

Σ ",, r ι

Here, fine constant and α are a coefficient. The time sequence of the normalized vectors u \ of characteristic parameters, which are obtained in this way, is stored in the sample buffer memory 27 under the control of the control circuit 21. In the micromusler buffer memory 28gc of FIG

Vector of characteristic parameters V _k (k : = 1 -> n) of η normalized languages stored, so-called micro-patterns. such as vowels which have characteristic features and have been obtained with the same processing as above. Here l \ need not be a time sequence of characteristic parameters of the vector. In other words, since the pronunciation of vowels, for example, is considerably stable. the vector of the characteristic parameters can be sufficiently representative at a point in time within this stable period. Fig. 4 shows the vectors of the normalized characteristic parameters of 5 vowels uttered by an adult male. In this embodiment, five micro-patterns corresponding to the five vowels are used.

The distance calculating circuit 29 reads the vectors of the characteristic parameters stored in the pattern buffer memory 27 and the micro-pattern buffer memory 28 by the control action of the control circuit 21, and calculates the distances between the vectors of the 16 normalized characteristic parameters u ' _t . which are stored in the pattern buffer memory 27, and the vectors of the 16 normalized characteristic parameters Y \ (k = 1 ^ 5) of the five micro-patterns. Finally, five distance parameters v * (A = 1 ^ 5 and ι a specific point in time) are stored in the parameter buffer memory 30. In this case, the distance parameter .vf is obtained by the normal distance calculation, which can be expressed by the following equation:

vf = Σ Kl ", ^k > ^2. it = 1-5).

Thus, the time sequence becomes five distance parameters . * J stored in the parameterization buffer memory 30. This time sequence of parameters is transmitted to the must-adapting unit 3 shown in FIG subjected to an adjustment operation with the parameters of the standard pattern for which the data already are compressed and registered in the same micro-pattern. Thus, the result of the recognition is used as the output signal! the isolation unit 4 in FIG.

In the system of conventional speech recognition devices, the data in the buffer memory is stored 25 stored time sequence of 16 parameters for the characteristic to be supplied to the pattern adaptation unit 3 Parameters have been used, while in the method according to the invention the time sequence of only five parameters can be used.

In general, the vectors of characteristic parameters u \ or \ ' ^k consist of the same number of bits. eg 16 bits as the number of bits of the distance parameter vf. Accordingly, according to the invention, the number of bits of the characteristic parameters required for the pattern matching in the pattern matching unit 3 can be reduced to 5 16 compared with the conventional method, and thus the time required for the comparison operation with the characteristic parameters of the Shorten standard samples by the same amount. Thus, the memory capacity for storing standard patterns can be reduced.

In the method according to the invention, it is necessary carry out the comparison operations with the characteristic parameters of the micro-patterns separately, however, the number of micro-musicians is at most about 5 and is much smaller compared to the number of standard pattern ii. /.U. K). Thus, the comparison option can be virtually neglected with regard to the micromusier. In addition, experimental

■; Received data showing that even if the Number of bits of the characteristic parameters for the above pattern matching is decreased, deterioration of the recognition ratio can hardly be observed.

FIG. 5 shows a block diagram of an exemplary embodiment for the structure of normalization circuit 26. Mnster buffer stores 27 and 28 as well as distance calculation sound 29 according to FIG. 2.

In the Normicrungssehaluing 26 denotes the loading

i </ ugszeichen 261 an adder with parallel inputs, while the reference characters 262-1 to 262-16 denote divisors. The micro-pattern buffer 28 consists of five micro-pattern memories 281 to 285 Distance calculation circuit 29 from five micro-pattern

; ii Comparators 291 to 295 exist.

If, with such a structure ρ, characteristic parameters are read out in parallel from the intermediate buffer memory 25 at a predetermined time and applied to the normalization circuit 26, these will be

: <Input signal from adder 261 of the normalization

| f>

circuit 26 adds, which results in V (/ ".

The addition results obtained in this way are applied to the divisors 2o2-1 to 262-16, with which a division

<i) is carried out according to the above equation (1) between these input signals or the characteristic parameters u " (_ / = 1 ~ 16) applied as input signals. After the normalization, the result is temporarily stored in the sample buffer memory 27.

Next, they will be excluded from the pattern buffer memory 27. normalized ρ characteristic parameters parallel to one another to the micro-pattern comparators 29i to 295 angclogi. On the other hand, since the micro-pattern stores corresponding to the comparators

There are 281 to 285 extracted 16 characteristic Parameters are applied to the comparators 291 to 295, both of the characteristic parameters of Input signals are compared with each other and their distances are calculated in these comparators. Demented

The outputs of the micro-pattern comparators become J> speaking 291 to 295 at a time sequence of 5 parameters and stored in the parameter memory 30. Using the parameters stored in this way in the parameter register memory 30, a Adaptation to the standard pattern carried out in the pattern adaptation unit 3 according to FIG The recognition result is used as the output of the decision unit 4 generated.

For the sake of completeness, this may be pointed out

<5, that the circuit arrangement according to FIG is also controlled by the control circuit 21 in FIG.

Although 16 dividers 262-1 through 262-16 are shown in FIG are, it is also possible to provide only one expensive and

») Normalize each characteristic parameter on a time divider basis using the one divider. In addition, the micro-pattern memories 281-285 can be used in common with the micro-pattern comparators 291-295.

It is also possible to provide such a circuit, soft the time sequence of the five characteristic ones obtained from the specified micro-pattern comparators Parameter smooths and the parameters again

aile samples one or two intervals to get the characteristic Parameter in the direction of your Zeil axis expands to compress I 2 or 13. According to the test results, there is hardly any impairment of the recognition ratio / u when using these compressed parameters.

Furthermore, it is also possible to carry out a scan at a point as a representative point, where i a change with time becomes maximum after the time series of 5 characteristic parameters is smoothed.

In the embodiment of Fig. 2 are different Memories 23, 25, 27, 28 and 30 are arranged separately, these memories being shared by a common memory can be replaced. υ

Although the output signals of the filter bank are used as the parameters representing the characteristics of the speech input signal in the embodiment described above, the recognition of the speech input signal can also be realized by means of other parameters m , e.g. B. with linear prediction coefficients, partial autocorrelation coefficients or the like.

In addition to the characteristic parameters of the vowels, other characteristic parameters such as: 5 eg frictional consonants (eg "S" or the like) or nasal consonants (eg "N" or "M") can be used as micro-patterns either alone or in combination with the vowels will. It is also possible to use parameters that are artificially specified and have intentionally emphasized characteristics.

In the above embodiment, the micro-pattern driving circuit calculates the distance, but the probability can also be calculated. In each Trap, any method can be used as long as comparison results of two samples are shown will.

Although five micro-patterns are used in the embodiment described above, can The aim of the invention can also be found using «1 of at least one micro-pattern without being specifically limited to the above embodiment is present.

From the above description, when carrying out the comparison between ρ characteristic parameters of the speech input signal and ρ characteristic parameters of η micro-patterns in the method according to the invention, the number of characteristic parameters to be compared with one another in the adaptation unit compared to the prior art is n / p (with the proviso that n <p ) is compressed. Since the processing time in the pattern matching unit and the decision unit can be shortened considerably in this way, it is possible to realize rapid recognition, namely recognition in a real-time arrangement or else to increase the number of words to be recognized. The memory capacity for storing the standard patterns can also be greatly reduced.

Thus, in the method according to the invention, the characteristic parameters of the speech input signal effectively compressed to improve the recognition speed and the recognition processing amount increase.

The method for speech recognition described above is thus characterized in that the comparison operations between ρ characteristic parameters.

which are calculated by calculating the voice input signals to corresponding / express points, and the respective /) characteristic parameters of / 1 given, specific languages, carried out to obtain characteristic parameters / u. Then it takes place a sort in terms of coincidence, between the resulting characteristic parameters and characteristic parameters of standard language, in the same way by performing comparison operations with characteristic parameters of specific languages have been obtained in order to do the speech recognition in this way.

For this purpose 4 sheets of drawings

Claims (1)

Patent claims: 1. Procedure for speech recognition in which from Speech input signals are determined with characteristic parameters s and with characteristic parameters are compared to those for given standard languages have been determined and saved, characterized by the following process steps: n>